TIO-Former: Ultra-Lightweight 6-Directional ToF-Inertial Odometry for Nano-UAVs

· Editorial Team estimated
nano-uav odometry tof-sensors inertial-fusion edge-computing streaming-transformer

Autonomous nano-UAV navigation demands accurate ego-motion estimation under severe size, weight, power and computing constraints, yet cameras and LiDAR exceed payload limits, optical flow degrades in low-texture scenes, and inertial-only estimation drifts. Multi-zone time-of-flight arrays are a lightweight metric complement, but recovering 6-DoF motion from only 384 ranges per frame must contend with invalid returns, anisotropic observability and computation that scales with flight time. TIO-Former is a camera-free, optical-flow-free and mapless range-inertial odometry framework driven by an IMU and a 15 g payload of six orthogonal 8x8 ToF arrays. Its frontend pairs consecutive range grids with a bilateral gated difference, while IMU-guided cross-attention routes directional features conditioned on platform kinematics; a streaming causal transformer couples an uncompressed local KV cache with compressed Chunk-FIFO memory so inference cost and memory stay bounded regardless of flight duration. In real flights it cuts open-loop position error by 54.4% versus nano-UAV optical flow and by 66.4%-89.1% versus learned inertial baselines, running on an edge RISC-V companion at 10.466 ms P95 latency and 6.324 MiB peak resident memory. Code is released.

Paper · arXiv:2609.17198

Ego-motion estimation for autonomous nano-UAVs has long been squeezed by SWaP-C limits: cameras and LiDAR exceed the payload budget, optical flow degrades in low-texture scenes, and inertial-only estimation drifts over time. Multi-zone time-of-flight (ToF) arrays look like the natural complement, being light and metrically direct, but recovering a 6-DoF pose from as few as 384 ranges per frame means simultaneously handling invalid returns, anisotropic observability, and computation that grows with flight duration.

Core Innovation

TIO-Former is a camera-free, optical-flow-free and mapless range-inertial odometry framework whose only inputs are an IMU and six orthogonal 8x8 ToF arrays totaling 15 g. Three design choices carry the work:

  1. Frontend range alignment: a bilateral gated difference pairs consecutive range grids, suppressing invalid returns and noise;
  2. Kinematics-conditioned feature routing: IMU-guided cross-attention dynamically selects directional features according to the platform’s current kinematics, easing the uneven observability across the six arrays;
  3. Bounded-memory streaming inference: a streaming causal transformer couples an uncompressed local KV cache with a compressed Chunk-FIFO memory, so per-frame inference cost and memory footprint do not grow with flight duration.

Results

  • Real flights: open-loop position error falls 54.4% relative to nano-UAV optical flow and 66.4%-89.1% relative to learned inertial baselines.
  • Environment and degradation robustness: evaluated across multiple environments and under severe sensing degradation.
  • Edge deployment: on an edge RISC-V companion computer, 10.466 ms P95 latency and 6.324 MiB peak resident memory, under 5% of system RAM.
  • Code released at https://github.com/Ly041021/TIO-Former

Limitations

The headline metric is a relative reduction in open-loop position error; the abstract gives no absolute error magnitude, flight distance or speed profile, so the task envelope where this is usable remains unclear. Baselines cover optical flow and learned inertial methods but not camera-inertial odometry, the very thing the design aims to displace. The range and field-of-view limits of the six ToF arrays, and their behavior under strong sunlight or highly reflective surfaces, are not discussed. The 10.47 ms figure is a P95 latency and it is not stated whether the rest of the control loop is included.

Industry Implications

The practical message is trading sensor choice against algorithmic burden: if low-cost, low-precision ToF arrays plus a suitable fusion network suffice for 6-DoF estimation, nano-scale platforms no longer have to pay the weight and power cost of visual odometry. For indoor inspection, confined-space search, and flight through pipes or crevices where the payload budget is measured in grams, this is a deployable alternative. The bounded inference and memory property may matter more than the accuracy numbers for long-duration missions, since it decides whether the system runs for hours on a fixed-memory MCU-class chip or overflows after a few minutes. Released code lowers the replication barrier.